Your Disk Is Not Cloud Storage
self-hosted = disk + power + a 2nd disk the number that matters here
By Stan K.
A QLC SSD that stalls under mixed load, a ZFS array that eats half your RAM, and the honest math of self-hosted storage: electricity, redundancy, and a backup for the backup. When $1.35/month on R2 beats the disk you already own.
Every storage conversation eventually reaches “why not just use the disk I already have?” We run self-hosted storage, and we’re the right people to answer — because we’ve watched ours fail in the specific ways self-hosted storage fails.
The QLC lesson
Consumer SSDs are mostly QLC now, and QLC has a failure mode nobody prints on the box: under sustained mixed load — simultaneous reads and writes, which is exactly what a “storage server that also does work” produces — throughput can collapse to hard-drive-worse speeds while health metrics stay green. We spent weeks diagnosing stalls that looked like GPU problems, looked like RAM problems, and were neither. The disk was simply drowning. Zero out-of-memory events, plenty of free RAM — and a chronically saturated consumer SSD underneath everything.
The lesson generalizes: a disk in your machine is a component of your workstation, not a storage service. It shares a power supply, a bus, an OS and a failure domain with everything else you do.
The ZFS tax
The “serious” self-hosted answer is ZFS — and ZFS is genuinely excellent. But know the price:
- The ARC cache defaults to up to ~50% of system RAM (tunable via
zfs_arc_max, which you will want to do on any shared machine) - Deduplication, the famous ZFS feature, is widely advised against — its RAM cost exceeds its value on most datasets
- ECC RAM is recommended, which drags the whole platform cost up
- Every pool wants a second pool, or at least a second box, for the “self-hosted redundancy” claim to be true
Self-hosted storage is not one disk. It’s disk + redundancy + monitoring + the discipline to test restores.
The honest math
What $1.35/month buys on Cloudflare R2 (100 GB, unlimited egress, 11 nines of durability, zero ops) versus self-hosting the same:
| Cloud (R2) | Self-hosted | |
|---|---|---|
| Media | $0 | disk $60–120 + a second one for redundancy |
| Running cost | $1.35/month | electricity ~$3–8/month for an always-on box |
| Failure handling | provider’s problem | your evening, your restore drill |
| Fire/flood/theft | survived | survived only with off-site copy — see below |
| Access from anywhere | built-in | your network config, your uptime |
The crossover is real but higher than people expect: self-hosting wins on bulk, private, read-rarely data — 10 TB of media and archives costs $69.50/month on B2 but ~$200 once on disks. That’s the Wasabi/B2 archive territory, and it’s legitimate.
A self-hosted copy with no off-site copy is not redundancy — it’s a single point of failure with extra steps. The 3-2-1 rule exists because the building burns down, not because the disk fails. Which means real self-hosted storage always ends up paying for cloud anyway — as the backup layer. Where those backup costs hide is its own article.
The synthesis we actually run
- Work-in-progress, hot data — the machine’s own disk, because latency matters and loss is tolerable
- Served files, anything a user downloads — zero-egress cloud (R2 or B2+Cloudflare), because serving from home is an uptime hobby
- Bulk archives — cheap cold cloud ($6.95/TB class) or local disks with an off-site story
- Backups of the above — separate tools, separate targets, append-only against ransomware
That layered picture — not “cloud vs local” — is the honest answer. The full provider pricing lives in the Cloud storage hub; the self-hosted software options (SeaweedFS, Garage, and the MinIO licensing situation) are in its second table.